Improving the Accuracy of Multispectral Image Clustering by Means of a New Initializing Method

Abstract

One of the disadvantages of the k-means clustering method lies in its dependence on the position of initial centers. Various methods presented so far, have emphasized the use of distance and density of points in the feature space as the most important features of the initial centers of clusters. In the method proposed in this work, the density of points, distance of centers from each other and their distribution in the feature space have been taken into consideration. The method acts on the base of a sub-set of image data and results in defining a standard using the distance of points from the centers of clusters, the density of points and the standard deviation of the distance of the points to the centers of clusters. The initial centers are then selected reiteratively. In this research, the k-means method has been used for clustering the multispectral imagery. The algorithm has been applied on simulated images and real satellite multispectral data. The results were compared with the reference maps. Some prevailing methods for the selection of the initial centers have been implemented for comparative purposes. The results showed that the proposed method, as compared with the existing methods, not only increases the accuracy, but also decreases the number of necessary iterations for clustering, at the same time maintaining the robustness of the method.